Our paper on stochastic ReLU functions for private inference will appear at Neural Information Processing Systems (NeuRIPS) 2021! The work is led by EnSuRe alum Zahra Ghodsi, and in collaboration with Brandon Reagen and Nandan Jha. The latency of many cryptographic private inference schemes is dominated by ReLUs. Circa introduces a new stochastic ReLU function that occasionally outputs incorrect values (for example, allowing small negative values to pass through to the output), but has up to 5x lower private inference latency. We show that SoTA DNNs are robust to these occasional errors and only incur a small accuracy drop.
